One 3D Object That Becomes Two Different Things When You Turn It

Carve a single statue so that from the front it's a rabbit, but spin it 90 degrees and it's a teapot. That's a 3D visual illusion: one solid object holding two completely different shapes depending on where you stand. JanusMesh builds these from text prompts in 3 to 5 minutes, without any training first.

The hard part is making the two shapes fuse into one believable object instead of two halves crudely glued together. Older methods either ran slow and produced garish, oversaturated colors, or stitched shapes so clumsily you saw obvious seams and one shape's features bleeding into the other.

JanusMesh splits the work in two. First it grows the geometry. It generates each target shape in parallel and merges them using a Signed Distance Field — a map that, for every point in space, records how far you are from the object's surface. Blending these maps lets the two shapes melt together smoothly, with a vision system (CLIP) checking that each shape still reads correctly from its own viewing angle. Second, once the merged form exists, it paints the surface — projecting view-specific image detail onto the geometry so each angle looks right.

The paper reports its method beats existing ones on shape integrity, recognizability, and speed, but those comparisons rest on the authors' own experiments, not an independent benchmark.

Why you should care: Illusion sculptures that took slow optimization and came out color-blotched now take a few minutes and look realistic — putting custom dual-view 3D objects within reach of anyone who can type a prompt.